Can AI slay the prior authorization dragon?

Heather Bassett, M.D., CMO of Xsolis

The insurance industry claims that prior authorizations are needed to ensure care quality and eliminate wasteful spending, but any physician will tell you they are mostly a waste of resources that takes time away from patients. Many of the tasks required for prior authorizations — or appealing their denials — could easily be automated, and many companies have done exactly that. But how much can technology help? Will artificial intelligence (AI) make prior authorizations obsolete? Or will payer use of AI just create more problems?

Medical Economics spoke with Heather Bassett, MD, chief medical officer of Xsolis, to find out how technology is affecting the prior authorization process. The transcript has been edited for length and clarity.

Medical Economics: Can AI help reduce these administrative burdens and make the prior authorization process better?

Bassett: AI is our best opportunity to start to move the needle in this space. The advantage of AI is you not only have your predictive machine learning models that can help you understand whether a case is likely to be approved or not, but you’re [also now] starting to see generative AI. Everyone has heard of ChatGPT, which is taking things to the next level of being able to summarize clinical information.

I think we’re at the precipice of seeing quite a bit of automation in this space, which ultimately will benefit not only physicians but also the patient, as well as provide better outcomes for the patient and offer more transparency and health care that is less confusing.

Medical Economics: Are we going to get to a point where it’s just the payers’ use of AI canceling out the physicians’ use of AI in prior authorizations?

Bassett: One area where we’re seeing generative AI expand is creating appeal letters, as it’s very good at summarizing clinical information. So, you wonder if the physician or the hospital is using generative AI to create an appeal letter, and then the rebuttal coming from the payer is, too. Are we going to enter a battle of the bots and those types of scenarios?

But if you take a step back, there is a lot of volume. I think it was 50 million prior authorizations in 2023, or, if you speak to physicians, an average physician does 39 prior authorizations a week. And it takes roughly 13 hours between the physician and their staff to take care of those authorizations. Of those authorizations, roughly 7% to 10% are denied, so that does mean about 90% are approved. It’s a very manual process to get those 90% approved. So, if you’re able to use AI to tackle that portion, regardless of if there are some slight increases in denials or other things people are worried about, you’re going to see tremendous benefit from a physician standpoint and from a hospital standpoint in automating the large number of prior authorizations that are approved.

Medical Economics: How much time do you think an automated prior authorization process could save physicians, and what challenges lie ahead?

Bassett: There are some numbers out there that say you could save 70% to 80% of the time that physicians spend on prior authorizations, so there’s definitely an opportunity for a big win. Part of the reason I ended up working for Xsolis is I was really starting to lose the joy of practice because there are so many administrative layers, whether it be prior authorization or other things that physicians end up having to do that takes them further away from the patient. The whole reason we went to medical school was to take care of the patient, and that’s a huge win to help solve that.

But the counter to that is the adoption of new technologies; it’s hard. Physicians can be a little stubborn at times. You kind of develop what you consider a very good process to be as efficient as possible. Getting that change management piece — getting them to be more efficient with new technology — can be hard.

Another big piece that we’re seeing around AI is trust. How can I trust that this is going to do what you say it’s going to do, and that it’s not going to create more work for me because of errors? You have to be considered a responsible company with responsible AI. When you think about innovation, it also creates an environment that allows the physician to gain trust — or whoever the end user is — by understanding the processes that are being put in place.

Medical Economics: What do you think prior authorization will look like in five to 10 years from now?

Bassett: I think things are just inherently going to change over the next couple years. It’s tricky in the five-year space. I am concerned about some of the changes that the new administration is putting in place.

My concern is there’s going to be a shift, and, unfortunately, I feel like health care may end up having more pressure put on it. And when you start to put pressure on the health care space, the innovation piece tends to be moved to the side.

We have seen a lot of investment in innovation by health care organizations out there — more so than we’ve seen in years — and it’s because they recognize that innovation. AI is how we’re going to fix problems like prior authorization. We’re going to fix problems around clinician burnout. Staffing challenges are real and have the potential to decrease quality outcomes that patients expect from the health care system. I do see the generative AI piece expanding. AI is very good at summarizing clinical information, which can help with crafting the information you send to the payer to get the prior authorization and help with appeal letters. Then if you go to the next step, we’re hearing a lot about ambient AI, and it’s really in the environment, passively taking information and then adding context to it. Primary care can bring a phone in and record the entire conversation, [but] it’s not just recording it; it’s taking that information [and] consolidating it into a note that the physician can review. It’s already saving physicians large amounts of time. Now, take that and think about agentic AI, where you are allowing it to do some tasks autonomously in order to get to a goal.

Think of prior authorization. For example, I’m a primary care physician, and I’m talking with my patient who’s going to need this medication or procedure. The AI can not only record the information but start to pull information to round out that prior authorization and even kick it off. Once we start to take all these new technologies and put them together, I think there’s tremendous opportunity.

Medical Economics: Most doctors will tell you that the electronic health record [EHR] rollout was a disaster because there weren’t enough doctors involved in the development process. What can physicians do to influence the development of these prior authorization technologies to make sure they better serve both their needs and their patients’ needs?

Bassett: That’s a great example. I was at an AI conference a couple years ago, and somebody from Google’s health care arm mentioned that exact thing. This is really when ChatGPT hit the stage in 2023. Basically, it was a call to action to the physicians in the room.

Remember the EHR? We did not have a voice. It basically was a glorified billing tool when it first rolled out. They’re better now, but it was [a] reminder [that] this is our time to have input on what the next iteration looks like. I think education is the first piece — really understanding what AI can do and what its limitations are. You don’t need to be a data scientist, but I think education is key.

There are a lot of organizations, like the American Medical Association, that are working with Congress to ensure that there are processes in place to standardize and improve the prior authorization process, so get involved with those organizations. If you hear that companies are looking for physicians to be involved in projects, be willing to be part of that because that’s the only way your voice is going to be heard.

To be honest, if you’re an AI company or a health care system that is using your own resources to build something, you can build the greatest AI models since sliced bread. But if you don’t think about workflow or about how the end user uses it and what their day-to-day world is like, you’re going to have an unsuccessful rollout. So, it’s kind of on both sides.